A Deep Learning Analysis on Training Issues of Turmeric Leaf Early Disease Detection

In the current global climate change scenario, we prefer technology-based solutions for the highest crop yielding rate. Current advanced technocratic approaches such as Artificial Intelligence and IoT are two popular areas that help traditional agriculture practices for food security and health informatics. Especially agriculture-based countries like India, China, the USA, and Brazil need efficient, smart mechanisms for sustainable food and medicinal production. Promoting health and well-being is identified as one of the seventeen global goals of the 2030 Agenda for Sustainable Development [1], as stated in Sustainable Development Goal (SDG) 3 of the United Nations. SDG #3. which pertains to health aims to establish universal health and well-being [2]. This includes a resolute dedication to eliminating epidemics of infectious diseases such as AIDS, tuberculosis, and malaria by the year 2030. Furthermore, it strives to establish Universal Health Coverage (UCH) and guarantee universal access to safe and efficacious vaccines and medications [3].

It is essential to support vaccine research and development and provide affordable medication access for this process to function. To achieve SDG3, currently, there is a need to introduce advanced technological practices for higher medicinal crop production. So far, many researchers and technocrats invented smart technology for easy agriculture practices such as precision agriculture, Crop Health Analysis, Climate-smart advisories, Yield forecast, Crop Damage Assessment, Supply Chain Management, Seed Production, Agro-Big Data Analytics, and smart farming practices. In addition to natural disasters and irregular weather patterns, medicinal plant diseases are another major cause of insufficient medicine production in the world [4]. The World Health Organization reports that 80% of the global population uses botanical remedies to fulfill at least a portion of their fundamental healthcare needs. Approximately 21,000 plant species have the potential to be utilized in medicinal applications, per the WHO [5]. It is a known fact that India and China largely depend on herbal treatment. Recently the Indian Govt. has proposed the concept of ‘Integrated Medicine’, where the Allopathic and AYUSH departments will work together to provide the best treatment that can be offered in one place [6]. At present the world is moving towards the usage of herbal medicines due to the less side-effect. Moreover, herbal medicines have become very popular in recent times due to wide usage in the treatment of COVID-19 (SARS-CoV-2). Especially the usage of Phyllanthus emblica, solanum indicum, Solanum surattense, Terminalia bellirica, Ficus religiosa, Piper longum, Curcuma longa Alhagi camelorum etc., are a few herbal medicinal plant species that are very frequent in COVID-19 treatment [7]. Among all other medicinal plants, Curcuma longa (Turmeric) is one of the herbal medicines that is used extensively in antibiotic preparation [8]. Cancer, rheumatoid arthritis, dermatitis, skin cancer, wound healing, urinary tract infections, and liver diseases are just some of the things that it can help with, Curcuma longa (Turmeric) is also employed in the treatment of these conditions [9, 10]. India is the leading producer and exporter of Curcuma longa (Turmeric) globally and shares 86% of global production. Among the Indian states, the state of Telangana stands first place for the Curcuma longa (Turmeric) [11, 12]. The state of Andhra Pradesh is one of the foremost contributors to the overall production of Curcuma longa (Turmeric) in the country after Telangana state. Among the 24 finest varieties in the countries, the Duggirala variety has a special purpose with rich medicinal values. Figure 1 shows the production and average price of the Duggirala variety in the state of Andhra Pradesh. The statistical results show the demand for Curcuma longa (Turmeric) is very high irrespective of the production. The production of Curcuma longa (Turmeric) is reduced due to COVID-19 on the global export and import of goods and unfavorable climatic conditions. This study focuses on the early detection of diseases in Curcuma longa leaves. This study aimed to determine the most effective Deep Learning (DL) training network that can achieve 100% accuracy when applied to the disease dataset of Curcuma longa (Turmeric- Duggirala variant). The collection consists of photographs depicting leaf blotch, Colletotrichum leaf spot, and Cercospora leaf spot. The deep learning mechanism is one of the most advanced technologies that can accurately identify items based on how they appear in photographs. Two stages are included in the deep learning approaches, which are training and validation. It is generally accepted that the impact of the correctness of the data training would be reflected in the validation metric, which is known as the mean correctness Precision (mAP). For the most part, when the accuracy of the training is lower, the validation accuracy scores are significantly lower. Therefore, achieving a training accuracy of one hundred percent is a challenging challenge for increased mAP. Typically, the performance of the training method is categorized into Gain and loss functions. When the Gain achieves 100% accuracy, it enables the classifier to identify objects with greater confidence levels. When working with smaller datasets that are balanced, it can be especially challenging to determine the appropriate values for training parameters, such as the learning rate and the number of epochs. This study also examined the most effective training optimizers for training the smaller diseased Duggirala Curcuma longa (Turmeric) leaf dataset, including Stochastic Gradient Descent (SGDM), Root Mean Squared Propagation (RMSProp), and Adaptive Moment Estimation (ADAM), along with vector distance (L2Norm, Global-L2Norm, and Absolute) to achieve higher levels of accuracy.

Figure 1. Duggirala turmeric variant production and average price during 2017-2023

Data Source: AP-Turmeric yard, Duggirala

The present study investigated the optimal Deep Learning (DL) training network for attaining 100% accuracy on the illness dataset of Curcuma longa (Turmeric-Duggirala variation). The collection comprises photos of leaf blotch, Colletotrichum leaf spot, and Cercospora leaf spot. Training parameters such as Learning Rate, Epochs, Gradient Threshold Method, etc., play a crucial role in object detection for deep learning architectures. In most of the cases, the dataset is segregated into 70% for training and 30% for testing. During classification, the performance of the algorithm is based on trained data by specifying a Gain function. The nature of datasets is always dynamic and it is very difficult for the researcher to identify the right parameters to train and classify the datasets. Moreover, choosing the right parametric values for higher classification training and classification is always a challenging task. The discussion in more depth can be found in the sections that follow.

Duggirala Curcuma longa (Turmeric) Dataset: A novel, custom-balanced Duggirala Curcuma longa (Turmeric) dataset collected 885 raw images that contain three classes of diseased leaves, such as leaf blotch, Colletotrichum leaf spot, and Cercospora leaf spot. Later dataset size was increased to 6,584 using pre-processing annotation techniques like image flip, rotation etc. and generated with balanced classes. The Duggirala Curcuma longa (Turmeric) Dataset is distinct and has not been previously available.

• Higher Training Accuracy: The results produced a higher training accuracy on smaller and balanced Duggirala Curcuma longa (Turmeric) Dataset with limited classes by tuning hyper parameters such as Learning rate, Epochs etc. This work achieved improved accuracy by systematically adjusting the hyperparameters of the optimizer and architecture.

• Impact of Optimizer on Training Accuracy: This work presents fresh findings on the influence of an optimizer on different widely-used deep learning approaches when applied to smaller, balanced datasets. This task is particularly tough, especially when dealing with damaged leaf data. This study specifically examines the performance of optimizers, including SGDM, RMSProp, and ADAM, in training neural networks. Our findings indicate that utilizing the RMSProp optimizer yields the highest classification accuracy among all designs while using lower learning rates and higher epochs.

• Impact of Vector Distance Methods on Classification Accuracy: This study presents a unique technical analysis by combining deep learning architectures with hyperparameters and optimizers on the Duggirala Curcuma longa (Turmeric) Dataset concerning the vector distance methods (L2Norm, Global-L2Norm, and Absolute) The study demonstrates the effectiveness of this approach by achieving higher accuracies on various common deep learning architectures and presenting multi-dimensional results. Based on vector distance methods the technical examination, defines that the ResNet-50 is the most suitable deep learning architecture for smaller datasets such as Duggirala Curcuma longa (Turmeric).

This work mostly concentrated on training-related concerns exclusively. Training the dataset is crucial for attaining superior test accuracy. The majority of research emphasizes end classification accuracy; however, if lesser test accuracy is achieved, a primary reason may be insufficient training accuracy resulting from suboptimal optimizer selection and hyperparameter configuration. Training accuracy is the paramount idea essential for achieving superior categorization in deep learning architectures. However, most research did not tackle the problem of inadequate training concerns. This paper examines training challenges with widely-used optimizers, including SGDM, ADAM, and RMSProp.

The current study focuses on conducting a comprehensive comparative analysis of deep learning architecture, optimizers, and hyper-parameters to determine the most effective training classification method, particularly for smaller datasets such as Duggirala Curcuma longa (Turmeric), which is distinct. These types of research are scarce in the data science literature, which typically examines a broad variety of characteristics and provides concise discussions. The current work just focuses on training accuracy and does not address validation accuracy in relation to the suggested dataset. Although the training accuracy hits 100%, there is uncertainty on whether the test accuracy can achieve 99%. The ultimate goal is to detect diseases in the leaf dataset; hence validation is a crucial aspect of the performance evaluation.

The subsequent sections of this study are structured as follows. Background research in the fields of agriculture and deep learning architectures is detailed in Section 2, early disease detection and training issues, and a novel image dataset of the Duggirala Curcuma longa (Turmeric) variant is prepared and discussed in Section 3. Similarly, Section 4 describes the mathematical notations and assumptions regarding explains details regarding methodology with the mathematical representation of gradient methods like RMSProp, ADAM, and L2 Regularization. Then section 5 describes the experimentation environment setup and section 6 explains the deep learning architecture. Then section 7 discusses the experiments of the models, Section 8 represents the results. Similarly, Section 9 detailed the performance evaluation metrics, then Section 10 represents the Anova test for the optimizers and then discussion. Lastly, the paper is summarized and suggests for best training classification methods in conclusions.

This study collects unique images of medicinal plant leaves of Curcuma longa of Duggirala variant. The Curcuma longa medicinal plant leaf dataset contains the Curcuma longa leaf blotch, the Colletotrichum leaf spot, and the Cercospora leaf spot. Three types of diseased leaves are gathered from the field, and the images of Curcuma longa are divided into three different classes. The Ayur Bharat and Deepherb is popular Indian medicinal plants databases which is open soruce This database establishes a comprehensive database encompassing 6,959 medicinal plants distributed across 28 states and 8 union territories, detailing their phytochemical properties and geographical distribution. It derives information from traditional knowledge, geographical indications, phytochemicals, and chemoinformatics.

The Ayur Bharat and Deepherb databases have limited turmeric images, and the Duggirala variation is unavailable. This study experiments on the innovative leave dataset obtained during fieldwork at the Crop area and Turmeric Board in Duggirala, Guntur District, Andhra Pradesh, India. HD 1024×768 pixels are used in the experiment. Clean the picture dataset for better categorization. Segmentation, picture scaling, contrast improvement, color correction, noise reduction, and feature extraction were used to preprocess the dataset.

To improve the classification analysis, the current Curcuma longa leaf dataset is analyzed using several color band indices that enhance ambient light conditions for color correction. Color band indices are commonly employed in crop image processing studies and are calculated by assessing the ratio of light intensity across different radiation bands. Light intensity is often averaged throughout the entire image or a designated segment of it.

The computation of the ratio of color band intensities functions as a normalization step to adjust for variations in ambient light conditions. This current Curcuma longa leaf picture dataset is analyzed using multiple Excess Red Index and Excess Green Minus Index (ERIEGMI) [23], Excess Blue color index (EBI), Excess Red Index (ERI) and Excess Green Index (EGI). ERI, EBI and EGI are improved the Classification analysis for Curcuma longa leaf images dataset samples.

Figure 2. (A) Original image; (B) Flip (Horizantal); (C) Rotate (-15˚); (D) Rotate (15˚); (E) Grayscale (15%); (F) Saturation (25%); (G) Brightness (15%); (H) Blur (Upto 2.5px); (I) Noise (Upto 0.1% of pixels)

Mainly three categories of sick leaves are included. Curcuma longa plant, popularly known as the Turmeric plant, is particularly subject to three diseases that greatly harm its rhizomes. The collection includes 6584 annotated photographs of leaf blotch, Colletotrichum leaf spot, and Cercospora leaf spot. The data is obtained using a high-resolution camera (Canon EOS 5DS R with 50.6 megapixels) from a specific area of land in Duggirala Mandal, Guntur District, Andhra Pradesh, India. Subsequently, the data undergoes picture pre-processing processes to create a dataset of superior quality. Pre-processing eliminates unwanted distortions and boosts important features necessary for the intended purpose. The noise data eliminated from the raw photos, categorized by class. In the production of the dataset, we utilize the following pre-processing techniques: (i) Data Profiling: The Curcuma longa dataset underwent profiling stages, which involved analyzing the distribution of colors, sizes, brightness, and shuffling. The open-source tool 'Data Gradients' is used to examine the Curcuma longa dataset. The analysis focuses on many criteria related to picture quality, such as convexity, fine features, segments, brightness, color distribution, aspect ratios, and resolution of leaf photographs. Figure 2 illustrates the Augmented techniques together with their parameters, while the distribution of three classes within the dataset, indicating that the dataset consists of 885 raw photographs classified into three categories, which were then increased to 6,585 images after annotation. (ii) Data Cleaning: We eliminated erroneous, corrupted, improperly formatted, redundant, or unfinished photos of Curcuma longa from the dataset. A total of 292 images of Leaf blotch were acquired, with 18 photographs being excluded from the collection. Similarly, I captured 305 shots of Colletotrichum leaf spot and then removed 29 of them. The majority of the discarded images were duplicates, while just a small number were found to be blurry. Following the same pattern as the previous instance, I gather a total of 288 photographs of Cercospora leaf spot. However, I exclude 15 of them as they possess inadequate resolution and brightness. We remove 62 anomalous data points from the original sample of 885, resulting in a balanced dataset of 885 photographs [24]. (iii) Balanced Dataset: Classifying balanced datasets yields much higher accuracy and reduces bias compared to imbalanced datasets. For the current research, a selection of 885 high-quality photos of pests was meticulously picked and resized to a resolution of 640 by 640 pixels in each category. Subsequently, these images undergo amplification [25]. (iv) Augmentation: Each pest class image undergoes a five-fold augmentation process, which involves rotation (four times), blurring (once), adding salt-pepper noise, and flipping (twice). This results in a total of 6584 augmented images across all three classes. The dataset is partitioned into three groups, allocating 20% for validation, 10% for testing, and 70% for training. The dataset distribution chart displays the distribution of Curcuma longa (Turmeric) leaf disease classes, which include three variants: Duggirala variation Leaf blotch (2192), Colletotrichum leaf spot (2208), and Cercospora leaf spot (2184).

Dataset

A total of 885 raw images of all classes were collected from the field. Among collected Colletotrichum leaf spot image took major portion (305) and followed by Leaf blotch (292), Cercospora leaf spot (288). A Data Training experimentation also conducted on imbalanced dataset and obtain 38 percent training accuracy for Squeeznet,29 percent for GoogLeNet and 42 percent for ResNet-50, when Hyper-parameters such as Learning rate (0.01), Epochs (25), Optimizers is SGDM and vector distance method is L2norm performs low accuracy with imbalance dataset. So, the rest of the work is discussed only on balanced datasets rather than imbalanced data set that resulted poor accuracy for all three training networks

Although we chose and adjusted the optimal hyperparameters based on a uniform probability distribution function (Epochs and Learning rate with equal intervals), the final findings were accepted using the Poisson distribution to suggest the most effective classifier.

5.1 SGDM

The SGDM algorithm exhibits the ability to oscillate in a direction that is most direct to the desired outcome. This oscillation can be mitigated through the incorporation of a momentum term into the parameter update. SGDM has been modified.

$\theta_{l+1}=\theta_l-\alpha \nabla E\left(\theta_l\right)+\gamma\left(\theta_l-\theta_{l-1}\right)$               (1)

The symbol γ represents the contribution to the current iteration from the previous gradient step. Momentum training is a method that may be used to determine this worth. Utilize the SGDM input option under the training options and use SGDM to train a neural network. The initial value for the learning rate should be provided using the initial learning rate training parameter. In addition, different layers and parameters may have different learning rates specified. While using SGDM with momentum, all parameters are learned at the same pace. Through the use of learning rates that change according to parameters and may automatically adapt to the loss function being improved, other optimization techniques seek to enhance network training.

5.2 RMSProp

The RMSProp calculates the moving average by considering the square of each parameter gradient and then averaging them.

$v_l=\beta_2 v_{l-1}+\left(1-\beta_2\right)\left[\nabla E\left(\theta_l\right)\right]^2$               (2)

The moving average decay rate is shown as β2. There are three standard decay rates: 0, 9, and 0,999. Averaging the squared gradients over durations of 10, 100 or 1/(1-β2) and 1000 parameter updates is required. By using the Squared Gradient Decay Factor (SGDF) training settings, you may pick β2. This moving average is used by the RMSProp method to normalize the updates of each parameter separately.

$\theta_{\iota+1}=\theta_\iota-\frac{\alpha \nabla E\left(\theta_\iota\right)}{\sqrt{v_l}+\epsilon}$               (3)

where, each part is divided separately. A small constant ɛ has been added to prevent division by zero and it successfully reduces learning rates for parameters with strong gradients while enhancing learning rates for values with mild gradients. While customization is possible via the Epsilon training option, the default configuration generally functions satisfactorily.

5.3 ADAM

The ADAM employs a parameter update method that bears a resemblance to RMSProp (whereas it is derived from ADAM) and incorporates a momentum factor. The parameter gradients and their squared values are monitored through the utilization of an element-wise moving average.

$m_l=\beta_1 m_{l-1}+\left(1-\beta_1\right) \nabla E\left(\theta_l\right)$               (4)

$v_l=\beta_2 v_{l-1}+\left(1-\beta_2\right)\left[\nabla E\left(\theta_l\right)\right]^2$               (5)

The Squared Gradient Decay Factor (SGDF) and Gradient Decay Factor training parameters, respectively, let you choose the β1 and β2 decay rates. Moving averages are used by ADAM to change the network's settings as necessary.

$\theta_{l+1}=\theta_l-\frac{a m_l}{\sqrt{v_l}+\epsilon}$               (6)

Parameter adjustments can gain momentum in one direction by employing a moving average of the gradient, provided that the gradients remain equivalent over multiple iterations. A significant portion of the gradient moving average is consumed by noise, leading to a corresponding decrease in parameter updates. One may specify through the utilization of the Epsilon training option. For some issues, a value as large as 1 is more effective than the default value, which is frequently adequate. If you wish to train a neural network using ADAM, the training choice for initial input is ADAM. In addition, the ADAM enhancement as a whole incorporates a method for reversing a bias that manifests itself during training. Utilizing the initial learn rate training parameter, configure the learning rate for every optimization technique. Learning rates that are suitable for a given optimization strategy are also dependent on how that rate influences that strategy. It is also possible to specify distinct learning rates for various layers and parameters.

5.4 Gradient clipping

The training process is characterized by instability and will deviate significantly after a few iterations if the magnitude of the gradient increases exponentially. A training loss that is NaN (Not a Number) or Inf (Infinity) indicates a "gradient explosion". Gradient pruning stabilizes training in the presence of anomalies and at higher learning rates, thereby mitigating the issue of gradient explosion. Gradient clipping enhances the efficiency of network training while preserving the correctness of the learned task. Norm-based gradient clipping adjusts the magnitude of the gradient while preserving its direction. The parameters 'l2norm' and 'global-l2norm' in the Gradient Threshold Method refer to gradient clipping techniques based on norms that partial derivative exceeding the threshold is subjected to value-based gradient cropping, leading to an arbitrary alteration in the gradient's direction. Although the behavior of value-based gradient clipping may appear counterintuitive, the network remains stable even when changes are of a negligible magnitude. The Gradient Threshold Method determines the absolute value of the gradient by employing a value-based gradient reduction strategy.

5.5 L2 regularization

An approach to mitigate overfitting involves incorporating a regularization term into the loss function E(θ) for the weights Eqs. (1)-(2). The weight decay term is also used in regularization terms. The form of the loss function including the regularization term is as follows:

$E_R(\theta)=E(\theta)+\lambda \Omega(w)$               (7)

The regularization function Ω(w) is denoted by where w signifies the weight vector and λ represents the regularization factor (coefficient).

$\Omega(w)=\frac{1}{2} w^T w$               (8)

7.1 Image data acquisition

Images of plant leaf diseases were acquired from the Customs Dataset, specifically from agricultural fields. The images were subsequently categorized into three distinct categories. The collection portrays Leaf Bloch and Spots, two prevalent leaf diseases that have the potential to affect the plant as mentioned above commodities. The diseases affecting turmeric leaves were selected for the compilation due to their global and Indian recognition, respectively. By default, each image is converted to a unique JPG file, which adheres to the RGB color space. A dataset has been collected from the field of research Area in Duggirala, Andhra Pradesh. The dataset contains three categories of disease leaf, which have leaf blotch, Colletotrichum leaf spot, and Cercospora leaf spot, respectively. The researcher collected private data from the field of Duggirala, Andhra Pradesh.

Adam is the suggested default algorithm and generally outperforms RMSProp and the mean epochs (25-50) and mean LR (0.01 to 1). Using the 'L2Norm' vector measurement method, ResNet-50 (SGDM) outperforms SqueezeNet (SGDM) by 48.8% and GoogLeNet by 24.55%. In ADAM usage, ResNet-50 (ADAM) has 36.74% higher training accuracy than SqueezeNet and 28.1% higher than GoogLeNet. Finally, ResNet-50 (RMSProp) outperforms SqueezeNet (37.13%) and GoogLeNet (40.88%). ResNet-50 (SGDM) outperforms SqueezeNet (SGDM) by 49.16.8% and GoogLeNet by 24.47% using the ‘Global_L2Norm’ vector measuring method. In ADAM usage, ResNet-50 (ADAM) has 38.29% higher training accuracy than SqueezeNet and 29.88% higher than GoogLeNet. The performance study of ResNet-50 (RMSProp) is 32.28% greater than SqueezeNet and 36.63% higher than GoogLeNet. ResNet-50 (SGDM) outperforms SqueezeNet (SGDM) by 49.32% and GoogLeNet by 32.50% using the ‘Absolute’ vector measuring approach. ResNet-50 (ADAM) has 37.03% higher training accuracy than SqueezeNet (ADAM) and 29% higher than ResNet-50 (SGDM) outperforms SqueezeNet (SGDM) by 49.32% and 32.50% using the ‘Absolute’ vector measuring approach. Tables 3 and 4 show training classification accuracy statistics.

The objective is very straightforward in classifying three Curcuma longa disease classes. We have created a synthetic Curcuma longa dataset that exhibits 100% linear separability, distinguishing points inside a circle from those outside. The Curcuma longa dataset contains a significant amount of clean, well-prepared data, which is proportionate to the complexity of the model. The signal in the dataset is potent and evident. The model's architecture is appropriately designed and not overly intricate for the task. A basic logistic regression model or a modest neural network may be the appropriate instrument. The model achieved 100% training accuracy, attributed to its simplicity, rigorous data verification, and high validation performance. The disease detection problem in the Curcuma longa leaf dataset is linearly separable, and the clean dataset removes noise and leakage. The model maintains 99.8% accuracy on the hold-out validation set and 99.5% on the final test set. In addition, the test results were subjected to an ANOVA test and achieved a higher test result of 98.7%. The SqueezeNet with ADAM optimizer is perfectly suited for a very simple, clean, and linearly separable problem. The task was easy, and the model learned it flawlessly.

The current study compares previous research, namely the work of Khan et al. [26], who conducted a comparative analysis of the performance of GoogLeNet, AlexNet, and SqueezeNet. By tweaking hyperparameters using electroencephalogram (EEG) data, they achieved accuracy rates of 94.99%, 94.61%, and 94.09%, respectively. The findings of the detection analysis indicate that AlexNet outperforms GoogLeNet and SqueezeNet. However, they concentrated solely on the epochs and learning rate, neglecting the importance of optimizers. The current study only fine-tunes the hyperparameters and also examines the influence of optimizers on training accuracy. Ullah et al. [27] specifically examined the performance of AlexNet, ResNet18, and SqueezeNet on a dataset consisting of 4333 photos belonging to eight distinct categories of road fractures. In this experiment, the training and testing images remained consistent throughout the epoch and iteration. The investigation's focus was not primarily on the choice of optimizer. The accuracy achieved with ResNet18 is just 85.2%. The proposed approach examined the GoogLeNet, SqueezeNet, and ResNet-50 models, with a primary focus on addressing training difficulties. We achieved this by fine-tuning hyperparameters, particularly emphasizing the usage of optimizers such as ADAM, RMSProp, and SGDM. Ashhar et al. [28] mostly looked into how well different deep learning models, such as GoogLeNet, SqueezeNet, DenseNet, ShuffleNet, and MobileNetV2, could classify lung tumours seen on a CT scan. They reached an accuracy of 94.53% using the GoogLeNet model. Their research did not take into account ResNet-50 and primarily concentrated solely on validation accuracy. Dahiya et al. [29] concentrated on training accuracy and used the Plant Village dataset, a dataset of 20,640 images representing 15 classes and 3 species: pepper, potato, and tomato. They applied this dataset to eight different deep learning architectures, namely AlexNet, GoogLeNet, MobileNet, ResNet 18, ResNet 50, ResNet 101, ShuffleNet, and SqueezeNet. Epochs, learning rate, mini batch size, and optimizer were the hyperparameters used. The range of epochs utilized varies from 30 to 50, with only the ADAM and SGDM optimizers being employed, while RMSProp is excluded. Out of the eight deep learning architectures, GoogLeNet demonstrates superior performance in accurately identifying larger datasets. Only two optimizers and a maximum of three epochs constrain their work. The current work applied three types of optimizers, namely L2Norm, Global-L2Norm, and Absolute, on six epochs ranging from 25 to 50. We set the batch size to 32 and employed vector distance methods. However, Dahiya et al. said that GoogLeNet is the most effective classifier. In our research, we found that ResNet-50 is the optimal model for classifying the smaller sick leaf dataset, mostly because of the variation in the vector distance model. In their study, Wagle and Harikrishnan [30] utilized various deep learning models, including AlexNet, VGG16, GoogLeNet, MobileNetv2, and SqueezeNet, to identify tomato leaf illnesses. They found that the VGG16 model exhibited superior accuracy and precision for most of the tomato leaf classes. However, it is worth noting that the study had limited scope for hyperparameter tuning. We compare the performance analysis with numerous cutting-edge research works. Finally, the current study is highly informative since it incorporates all hyperparameters to evaluate the performance of common deep learning algorithms on a small, balanced dataset [31-34]. This research is distinctive due to its integration of deep learning architecture with optimizer and distance vector approaches (L2Norm, Global-L2Norm, and Absolute) while also adjusting epochs and learning rates. This approach resulted in improved classification performance on a smaller dataset while maintaining minimal time complexity. This original technical analysis demonstrates the uniqueness of the research. The current effort focuses primarily on medical plants rather than commercial crops and traditional object identification. Medical plants have become particularly important in the aftermath of the COVID-19 pandemic. The current investigation is limited to three specific pre-trained models for the sake of simplicity. However, it is possible to further train the current dataset using additional deep learning models such as EfficeintNet, AlexNet, VGG16, DarkNet, PANet, ShuffleNet, NasNet Xception, MobileNet-v2, and others. This will allow for an evaluation of the effectiveness of the presented results. The current study is beneficial for researchers with smaller and balanced datasets, as they can achieve higher accuracy by appropriately adjusting the hyper-parameters.

Pandey et al. [35] Identifying diseases in medicinal plants is essential for the quality and effectiveness of herbal therapeutics. Improvements in morphological observation, histological assessment, molecular biology procedures, and imaging modalities have enhanced illness identification. Contemporary technology, including high-throughput DNA sequencing and real-time PCR, improves speed, precision, and efficacy. This highlights the significance of continuous efforts for sustainable production of medicinal plants and accessibility of herbal medicine. Their study specifically focused on DNA sequencing, but the current study focused on training issues in deep learning techniques to achieve higher test accuracies.

Pushpa et al. [36] research investigates three hybrid deep-learning models for the real-time identification of medicinal plant species. The models employ VGG 16, MobileNet, MobileNetV2, and ResNet50 as feature extractors. The final feature vector is derived by consolidating the extracted features. The hybrid model 3 surpasses other models, exhibiting enhanced performance attributable to feature channel rescaling. The models are lightweight CNNs designed for mobile applications and are anticipated to be augmented to encompass rare medicinal plant species for enhanced identification and biodiversity conservation. But their research not focused on SqueezeNet and GoogLeNet and also not focused on distance vector methods, which is an important parameter to achieve higher test classification.

Sharma and Vardhan [37], study seeks to enhance the precision and efficacy of identifying medicinal plants utilized in traditional medicine. A substantial dataset of medicinal plants and leaves is utilized to develop a deep learning architecture known as the Attention-based Enhanced Local and Global Features Network (AELGNet). The architecture identifies salient aspects from the images, deriving fundamental characteristics for both local and global extraction. The research demonstrated that AELGNet surpasses 14 existing methodologies, serving as an effective instrument for the precise and rapid identification of medicinal plants and leaves in both medical and industrial contexts. This work not considered the traning issues, which is plays a significant role in medicinal plant leaf disease detection that addressed in the present work [38, 39].

Achieving elevated training and validation accuracy through the selection of suitable optimizers and acceptable vector algorithms for classifying Curcuma longa leaf disease detection datasets present a continual challenge for academics and practitioners. This study primarily emphasizes datasets of medicinal plant leaves. This study achieved an optimal accuracy of 100% utilizing the ADAM optimizer with L2Norm distance vector, specifically for balanced datasets of Curcuma longa leaves. For example, additional entities such as automobiles, structures, and individuals may not yield superior outcomes based on the current optimizer's recommendations. This study identified the optimal combinations by adjusting different optimizers and vector distances to get higher accuracies for Curcuma longa-like leaf diseases.

We have created a synthetic Curcuma longa dataset that exhibits 100% linear separability, in identifying and characterizing medicinal plants for novel pharmaceuticals and therapies. It facilitates individualized care, conservation efforts, traditional medicine, and addresses taxonomic deficiencies by identifying and monitoring endangered species while safeguarding traditional knowledge.